Instruction-Based Self-Supervised Online Training of the Perceptual Subsystem of a Cognitive Robotic Architecture

Traditional AI systems often operate under the closed-world assumption, restricting their ability to
adapt in dynamic environments. We propose a cognitive architecture (CA) that expands its perceptual capabilities by generating object prototypes from user-provided natural language descriptions.
Each prototype is constructed using superellipsoid primitives, enabling structured and interpretable
shape representations. The CA employs these prototypes to train a convolutional parametric shape
encoder, using rendering parameterizations as automated ground-truth supervision. Once trained,
the CA employs the encoder to infer superellipsoid-based representations from real-world object
observations. A bidirectional mapping between superellipsoid parameters and natural language
terms allows the CA to translate inferred geometric features into human-understandable descriptions. We detail the design of the prototype representations, the synthetically supervised training
pipeline, and the language–geometry mapping process. Experimental results demonstrate that the
CA enhances its perceptual repertoire through our structured, interpretable object representations.